The mp_client crashes if run on Azure or any system where ethdev
ports are owned. In that case, the tx_buffer and tx_stats for the
real port were initialized correctly, but the wrong port was used.
For example if the server has Ports 3 and 5. Then calling
rte_eth_tx_buffer_flush on any other buffer will dereference null
because the tx buffer for that port was not allocated.
Also:
- the flush code is common enough that it should not be marked
unlikely
- combine conditions to reduce indentation
- avoid unnecessary if() if sent is zero.
Fixes:
e2366e74e029 ("examples: use buffered Tx")
Cc: stable@dpdk.org
Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>
Acked-by: Matan Azrad <matan@mellanox.com>
for (;;) {
uint16_t i, rx_pkts;
- uint16_t port;
rx_pkts = rte_ring_dequeue_burst(rx_ring, pkts,
PKT_READ_SIZE, NULL);
- if (unlikely(rx_pkts == 0)){
- if (need_flush)
- for (port = 0; port < ports->num_ports; port++) {
- sent = rte_eth_tx_buffer_flush(ports->id[port], client_id,
- tx_buffer[port]);
- if (unlikely(sent))
- tx_stats->tx[port] += sent;
- }
+ if (rx_pkts == 0 && need_flush) {
+ for (i = 0; i < ports->num_ports; i++) {
+ uint16_t port = ports->id[i];
+
+ sent = rte_eth_tx_buffer_flush(port,
+ client_id,
+ tx_buffer[port]);
+ tx_stats->tx[port] += sent;
+ }
need_flush = 0;
continue;
}